Cole Medin applies coding-agent practices to a production help desk in this 16-minute tutorial, made in collaboration with LangChain.
Improve the software around the model
- A harness is the software around a model that runs its tools, files and commands.
- Medin maps global rules to system prompts, subagents to delegation, skills to workflows and hooks to safety checks.
- LangChain’s Managed Deep Agents packages the instructions, skills and supporting code into a deployable agent.
- The roughly 30-line agent definition shown relies on that surrounding platform and its Markdown files.
Loop one: turn human corrections into memory
- When the agent surfaces uncertainty, route the case to a team inbox for review.
- Record the confirmed fix as a searchable lesson, with shared knowledge separated from each user’s private memory.
- Retrieve relevant cases when needed to keep the main instructions compact.
- In the demo, a member’s local changes are missing because Archon creates its worktree from
origin/main. - A human teaches the agent to push the commits and rerun; another member’s similar problem then receives the learned explanation.
Loop two: turn failures into instruction changes
- Use traces, records of the agent’s steps, to connect a failure with the conversation that caused it.
- LangSmith Engine proposes a concise change to the agent’s instructions for a person to inspect and approve.
- Keep version history so a change can be promoted or rolled back if it makes behavior worse.
- The workflow automates diagnosis and patch proposals while retaining human approval for changes.
Read the result as a small demonstration
- Medin reports moving from 4 correct answers out of 14 to 13 out of 14 after the learning loops.
- Caveat: these are results on his small demo question set, without a separate unseen test set described in the video.
- The uncertainty detection and private-memory separation are presented as platform behavior; the walkthrough supplies no independent reliability or isolation test.
“Harness engineering is not a one-time job.” — Cole Medin, 15:42